jtdean123 / immunohe

melanoma diagnostic H&E image transfer learning

immunoHE

Code to reproduce the work in "Prediction of TMB and an immune gene expression signature from melanoma diagnostic H&E images"

The code required to reproduce figures 3-5 in the manuscript is in the figures directory.

Instructions for reproducing all of the work in the manuscript from end to end:

Image Processing

Collated TCGA data

Collated TMB and immune signature data for TCGA-SKCM samples is found in data.tsv.

Download RNASeq data (optional)

First, install the GDC data transfer tool:
https://gdc.cancer.gov/access-data/gdc-data-transfer-tool

Then download the RNASeq data for TCGA-SKCM files as follows:

./gdc-client download -m gdc_manifest.2018-10-04.txt

Download and tile diagnostic H&E slides

1. Navigate to images/
2. Run the following to create a shell script for image pre-processing:

./make_tiling_jobs.py -s 512 -m gdc_manifest.2018-11-03.txt

This creates a script tile_jobs.sh that will coordinate downloading TCGA-SKCM diagnostic H&E slides, tiling, removal of the original slides, and selecting min(number of patches, 1000) patches from each slide.

3. Run this as follows:

./tile_jobs.sh

Normalize the slide patches

1. Navigate to images/
2. Run the following to create a shell script for image patch normalization:

./make_norm_jobs.py

This creates a script norm_jobs.sh that will normalize all of the patches using NORM.jpeg as a reference.

This also creates a tsv, image.paths.normed.tsv, that you will need for patch-label assignment.

3. Run this as follows:

./norm_jobs.sh

Feature Extraction

1. Assign TMB and immune signature patch labels by creating two tsv files images.paths.labels.immune.binary.tsv and images.paths.labels.tmb.binary.tsv.

./add_label.py

2. Use inception v3 to extract features from each patch.

./feature_extractor.py --samples images.paths.labels.tmb.binary.tsv --output features/tmb.inceptionv3.features
./feature_extractor.py --samples images.paths.labels.immune.binary.tsv --output features/immune.inceptionv3.features

This process takes ~8hours using a GPU and generates two 3Gb npy files.

3. Feature pooling

Pool the features that you just extracted as follows:

./pooler.sh

This will create max, average, and p-normed feature vectors for each slide and write them to the features/ directory. The pooled feature vectors used in this work are located in the features/ directory.